Description

This courselet examines whether qualitative information contained in Airbnb guest reviews provides useful signals for evaluating real estate offers and predicting listing prices.

The analysis uses Berlin listing and review data from Inside Airbnb, combined with geospatial information about nearby public transport, restaurants, supermarkets and schools from OpenStreetMap.

The workflow covers:

• Combining listing, review and geospatial datasets
• Cleaning review text and removing HTML, addresses and named entities
• Splitting reviews into sentences with distinct meanings
• Creating multilingual sentence embeddings
• Reducing embedding dimensions with UMAP
• Detecting semantic clusters with HDBSCAN
• Discovering recurring topics with BERTopic
• Generating readable topic labels with GPT-4o-mini
• Measuring topic-specific sentiment
• Converting review topics into structured listing-level variables
• Adding distances to nearby amenities
• Comparing different feature sets for Airbnb price prediction

The results show that using the 30 most informative review topics produces the lowest mean absolute error of €34.41. Adding topic-specific sentiment achieves the highest R² of 0.7108, while including all 150 topics does not improve prediction accuracy.

The courselet demonstrates how unstructured guest reviews can be transformed into interpretable quantitative features for real estate evaluation.